This manuscript presents a blue-band surface residual correction method for land aerosol optical depth (AOD) retrieval from DQ-1 WSI, combining machine-learning-based correction of the surface reflectance prior with a physically based LUT retrieval framework. The topic is relevant to AMT, and the reported results show a clear improvement over the static-prior retrieval. The manuscript is generally well structured, but several aspects of the methodology and validation require further clarification, particularly the role of MCD19A2 AOD in constructing the RF training target, the limited independent AERONET validation, and some details of the retrieval and post-processing procedures. These issues are important for assessing the robustness and reproducibility of the method, but do not appear to undermine the overall retrieval framework. I believe the manuscript can be considered for publication after the issues below are adequately addressed. My specific comments are listed below.

1. The RF model does not directly predict AOD, but the residual target used for RF training is constructed using MCD19A2 AOD. The manuscript should therefore distinguish more carefully between “not directly predicting AOD” and “being independent of external AOD information”. Some statements regarding the independence of the proposed approach from external AOD constraints appear somewhat strong and should be revised accordingly.
2. The independent AERONET validation contains only 27 station–date matchups. Although the reported improvement relative to the static-prior retrieval is substantial, this sample size remains relatively limited for assessing the overall robustness and general applicability of the method. The authors should state this limitation more explicitly and interpret the validation results with appropriate caution. If expanding the validation dataset is not feasible, the manuscript should also clarify the practical reasons for the limited number of matchups and more clearly define the conditions over which the current validation supports the retrieval performance.
3. Please clarify what is meant by “haze” in the cloud–haze screening procedure. Since aerosol haze is itself the target of AOD retrieval, it is important to explain how aerosol-loaded but otherwise valid observations are distinguished from contamination that should be masked. The authors should also clarify whether this screening could preferentially affect the retention of high-AOD cases.
4. Section 3.2: The manuscript considers both continental and urban aerosol models. Please indicate which aerosol model was selected for the main retrieval cases and clarify how the scene-level model selection was implemented. The selection criterion should be stated explicitly so that this part of the retrieval procedure can be reproduced.
5. In the comparison with MCD19A2, collocated pixels are excluded when AOD < 0.05 in either product. Please justify this threshold and report approximately how many samples are removed by this criterion. Since excluding low-AOD cases may influence the regression slope, bias, and other reported statistics, the possible effect of this filtering on the comparison should be clarified.
6. Section 4.1 / Table S3: The AERONET temporal matching window is relatively broad compared with the DQ-1 overpass time. Please justify the choice of this window and briefly discuss the possible influence of temporal aerosol variability on the validation statistics. Given the limited number of station–date matchups, it would also be useful to comment on the potential sensitivity of the reported performance to the temporal collocation criterion.
7. 2: The reference-date switch ratio is defined relative to the default value α = 0.06. Therefore, the minimum value at α = 0.06 follows, at least in part, from the definition itself and does not by itself provide independent evidence supporting the choice of this parameter. The corresponding interpretation should be revised, and the authors should clarify what additional criterion was used to justify α = 0.06.
8. (2): Please provide a short physical explanation for the brightness term in the reference-scene selection score. It is not immediately clear why a relatively brighter candidate surface state should provide a more reliable static surface reference. Since this term directly affects reference-scene selection, its physical or empirical basis should be made more explicit.
9. Section 3.3: AOD ≤5 is treated as a low-AOD condition for construction of the surface reference. This threshold appears relatively high, particularly for a retrieval that relies strongly on the blue band, where aerosol effects can remain substantial at such AOD levels. Please provide a clearer justification for this choice and discuss whether residual aerosol effects under these conditions could influence the constructed surface reference.
10. Section 3.2: Please provide some additional information on the implementation of the 6S simulations and LUT. In particular, please clarify whether the actual DQ-1 WSI spectral response functions were used rather than only the nominal central wavelengths. The manuscript also mentions an additional gaseous-absorption correction applied to the LUT; the implementation of this correction should be described in sufficient detail to allow the retrieval procedure to be reproduced.
11. Section 4.2: Please clarify whether the 28 April 2025 scene was used at any stage of hyperparameter tuning, feature selection, model selection, or other optimization. If it was used for any of these purposes, it would be more appropriate to describe it consistently as an internal validation scene rather than as a fully independent test case. This distinction is important when interpreting the reported generalization performance.
12. Several empirical thresholds are applied to the training residuals, predicted residuals, and corrected surface reflectance. Please explain the rationale for these thresholds more clearly. It would also be useful to report the fraction of retrievals affected by the clipping or bounding operations. If a non-negligible fraction of pixels is constrained, the possible influence of these empirical limits on the final AOD retrieval should be discussed.
13. The manuscript discusses the relative importance of the RF input variables. Please clarify how the feature importance shown in Fig. 3 was calculated. Since some of the predictors may be correlated, a brief note on how the importance values should be interpreted would also be helpful.
14. In the MCD19A2 comparison, please provide a quantitative and reproducible definition of the “residual cloud-contamination outliers” removed before calculating the cross-product statistics. The current description is somewhat subjective. The number or fraction of samples excluded by this procedure should also be reported, since outlier removal may affect the reported correlation, regression, and error statistics.
15. The very large number of collocated pixels in the MCD19A2 comparison should not be interpreted as the same number of statistically independent observations, because neighboring pixels within the same scene are spatially correlated. Please report the number of independent observation dates included in this analysis and make the distinction between pixel count and independent sampling more explicit. Date-by-date summary statistics in the Supplement would also be useful for assessing whether the reported agreement is consistent across scenes rather than being dominated by a subset of dates.
16. Section 4.1 / Fig. 4: Please provide the expected-error expression directly in the main text or figure caption. This would make the validation figure easier to interpret without requiring the reader to refer to the Supplement. It should also be made clear that this expected-error envelope is used as a reference benchmark rather than as a DQ-1-specific uncertainty specification.

I apologise for my delay in posting this review, due to some unexpected circumstances which arose since accepting the invitation.

This manuscript describes a method to obtain an a priori surface reflectance for aerosol optical depth (AOD) retrieval from the WSI satellite instrument over the Beijing–Tianjin–Hebei (BTH) region in China. It is a two-step approach, first creating a static reference data base (beforehand) and then applying a correction based on mismatch at 443 nm (on a scene-by-scene basis). The static data base is obtained using atmospheric correction of WSI data from low-AOD days using MODIS MAIAC AOD retrievals as an input. The correction step improves performance of the AOD retrieval, based on comparisons of results using the static vs. corrected surface reflectance against ground-based AERONET observations at one location. The idea seems promising and to my knowledge is novel. Generally a prior surface constraint is necessary because surface reflectance modeling errors are a large contributor to AOD retrieval uncertainty, particularly when AOD is low, and a single-view multispectral imager generally does not have sufficient information content to constrain both spectral aerosol and surface information from a single overpass. As such the manuscript falls within AMT’s scope. I appreciate the authors having incorporated suggestions made during the Quick Review process as this has made the manuscript easier to follow.

My main concern is about the extent of the data processed and analysed. It is very restricted: to the BTH region only and for three months only (March to May 2025). It is not clear how generalisable to other parts of the world or times the results would be (despite the authors’ claims that it is, it is not supported by the analysis). This short scope also makes it hard to judge the quantitative comparisons with AERONET. The data volume is not enough to make robust conclusions, and it is common for new algorithm papers to describe higher data volumes than this. Since the satellite launched in April 2022 it should be easy to process more data, even if only in the same study region. If nothing else this will also shed light on issues such as whether the initial 3-month data base is valid for other years, or times of year. Additionally, the aerosol optical models seem to have been selected for convenience and are not well justified by the analysis shown in the manuscript. These concerns are touched on by the authors in section 5.2 but I think more should be done on them in the current work. As a result, I recommend major revisions and encourage the authors to expand the analysis to make the results more robust. My specific comments are below:

1. Line 122: What is the satellite overpass time window? This should be stated here.
2. Line 126: A citation is needed for the MAIAC aerosol data set here. Probably: Lyapustin, A., Wang, Y., Korkin, S., and Huang, D.: MODIS Collection 6 MAIAC algorithm, Atmos. Meas. Tech., 11, 5741–5765, https://doi.org/10.5194/amt-11-5741-2018, 2018.
3. Line 136-140: I think everything from “To ensure” to “navigation data” should be deleted because it is to be expected that the algorithm is working in calibrated physical units and that geometry is needed. If this text is kept in some form then a link or reference to the CRESDA calibration data should be given (if this calibration is not something which is included by standard in the WSI level 1 data).
4. Lines 144, 150: Acronyms such as MNDWI, NDVI, NDSI should be defined at first use.
5. Sections 3.1, 3.1 and in general: the manuscript is not always clear about what it means by “reflectance”, e.g. line 170 talking about “surface reflectance terms” (what quantities?). I recommend adopting nomenclature from Schaepman-Strub et al (2006) where possible: https://www.sciencedirect.com/science/article/pii/S0034425706001167 Reading sections 3.1 and 3.2 talking about the use of MAIAC AODs and WSI reflectance, what exactly is being calculated for the static data base? The paper mentions Lambertian assumption using 6S for atmospheric correction but then talks about geometry being used for the random forest correction later. What I infer is being done is that the static data base is created using Lambertian atmospheric correction of WSI images using MAIAC AOD (at which wavelength, this needs to be stated?) but that aerosol optical properties (i.e. spectral and directional dependence of scattering and absorption) come from OPAC which is built into 6S (not from MAIAC – which creates an inconsistency). The multiple scenes used to generate the data base are combined to give the final data base value using rules based on brightness and NDVI (are these the “matching criterion” used to pick aerosol model mentioned in line 208? This is also unclear). Thus there is no direct accounting for BRDF effects in here. Then the random forest (RF) method used in the correction approach takes in geometry so I guess is making the first-order correction for BRDF effects (since if I read correctly it is trained on 3 of the individual scenes going into the data base), but there is no inherent physical BRDF model. Is all this interpretation correct? This all feels very empirical but I suppose if it works it works, at least within this limited domain. Although the authors state that the retrieval is not circular because the MAIAC AOD is not input into the process here, I am not fully convinced, because the training target of this RF method is a discrepancy between the data base and the surface reflectance obtained using MAIAC AOD so it is still indirectly in there. However, since the training was done based on low-AOD conditions it seems likely that the bulk of the RF correction targets e.g. BRDF effects and not indirectly consistency with MAIAC aerosols.
6. This does lead in to a general concern (touched on in section 5.2 but not enough in my view): the study presents results only for a small geographic area and for three months (March-May 2025). It is not clear how generalisable this will be. For example, I would expect changing vegetation phenology and changing solar angles throughout the year to influence apparent surface reflectivity which would likely mean that additional data bases are needed for other times of year (this is what e.g. the NASA Deep Blue approach does). It’s also not clear that a data base created for one year would be suitable for use in other years. Additionally, while this is a proof of concept for the BTH region, it’s not clear how the availability of suitable scenes to construct the static data base would be elsewhere in the world (due to e.g. availability of low-AOD, cloud-free, snow-free days and sufficient stability of surface characteristics). Presumably the WSI satellite is intended to eventually lead to AOD retrievals outside the BTH region (indeed, outside the cities themselves – the only maps shown are described as urban areas, we don’t actually see any broader maps) so I would have appreciated some more analysis (or at least discussion) of that. It is not unusual for new algorithm papers to have a somewhat broader spatial/temporal scope.
7. Line 195: the continental and urban OPAC models in 6S seem to have been picked because of their names and because they are convenient to use. There is no scientific justification provided. There should be some discussion of why this is a reasonable choice, or if it was because it was convenient, at least say that. These OPAC models are very old and I’m not sure how reasonable the urban one in particular is. Looking at the Supplement they have fairly similar spectral dependence (and there’s no evidence provided that this spectral dependence is common for the BTH region) and are quite absorbing (especially the urban model, with single scattering albedo, SSA, below 0.7 at all wavelengths). As one simple step the authors could look at e.g. AERONET retrievals in this region and see how closely these models fall to observed spectral dependence of AOD and SSA. If they are similar, this is evidence that it is a reasonable choice. If not, then more justification and/or a better model is needed.
8. Lines 201-203: More information is needed on the gas absorption cross sections and how the correction was done. Spectral responses are not shown, absorption cross sections are not shown, source of gas concentrations is not given. I’m assuming either the assumption of scattering-absorption decoupling was made or the cross-sections and spectral responses were built into 6S and it was handled there. But none of this is stated.
9. Table 2: the angular steps here feel really coarse, is this truly sufficient? Most other approaches use finer angular grids. Do the authors have an estimate of uncertainty introduced by this angular sampling? I would expect some uncertainty from the interpolation step because top of atmosphere (TOA) reflectance often does not vary linearly with angular changes.
10. Section 3.4: Just commenting to note that while I understand the intent of the correction approach I am not an expert on the RF method so would defer to other reviewers on whether the chosen method and level of analysis are sufficient for this part.
11. Line 302: the cost function used needs to be defined. I am guessing it is either absolute residual TOA reflectance, square residual TOA reflectance, or one of these normalised by some assumed measurement uncertainty? Since the thresholds mentioned later in this paragraph are small numbers my guess is probably not normalised?
12. Section 4.1: the data volume is really not adequate to make a meaningful assessment of performance. It is probably safe to say that the residual correction approach works better than the than the static LUT, but not much else. This is only 27 points from a single location, and no telling how representative that is of the broader region. The authors acknowledge this is limited but in my view agreeing it is limited is not enough, more data should be processed to get towards a more robust data volume. I would also suggest the authors do the AOD comparison at multiple wavelengths; since the spectral dependence of AOD is fixed in their retrieval approach to one of two models (continental or urban) it would be interesting to see if any one wavelength seems more stable than the others. Scatter plots could also be coloured by the AERONET Ångström exponent to see whether that seems to covary with the error of individual points. That would provide a useful indication of whether the remaining main error source is aerosol spectral dependence as opposed to something else (e.g. surface reflectance, SSA).
13. Section 4.4: For me, presenting Figure 9 as a trio of cutouts of municipal boundaries is not so useful without the broader regional context. My suggestion is just to show the whole broader BTH area these three locations all sit in (i.e. including the spaces between them). and add a panel showing a true colour image from this day so the reader had broader context. So this would be 5 larger panels instead of 12. About all I glean from the current image is that the residual-corrected method matches MAIAC better than the static one, but it’s hard to gauge much more.
14. Line 620: links and/or data DOIs are needed for this section.